CK-19 Atomic claim decomposition and scoring (FActScore; atomic fact checking)
Split generated text into atomic facts and score the share that a reliable knowledge source supports, and produce a supported proportion and a list of unsupported facts.
What it does
Splits generated text into atomic facts and scores the share that a reliable knowledge source supports. Artefact: the proportion of supported facts and a list of unsupported facts. Read the unsupported list.
When to use it
Use when a long model output contains many separate factual claims and you want a figure for how many are supported.
When it does not fit
Fits long outputs with many separable claims. Misfit: claims that are inferences and not facts stated in the source. The estimator uses a strong model, so the checker can share the generator's errors (inference).
Result format
The result lists the atomic facts split out of the generated text, each marked as supported or unsupported by the knowledge source. It gives the proportion of supported facts and ends with the list of unsupported facts.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text |
| Unit of analysis | Always applies | requires=claim |
| Appraises source credibility | Applies in some cases | no |
Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.
| Minimum material | a reliable knowledge source or the source documents, and outputs with several separable claims |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice for claims that are inferences and not facts stated in the source. The reported error rate came from one biography task and must not be transferred to other tasks. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | report |
| Kind of finding | contradiction; gap |
| Question answered | what |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | proceed with limit stated |
Library checks by stakes
Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: judgement dependent.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| decision-informing | spot check | optional | no |
| decision-informing | human review sample | needed; left out below high stakes | no |
| decision-informing | source-span citation | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | human review sample | needed | no |
| high-stakes | source-span citation | none | no |
| high-stakes | gold standard agreement | conditional | no |
| high-stakes | independent review | needed | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | Accepts a plausible claim that the source does not state. |
| Error sources | input data; language model |
| Certainty limits | Certainty drops when the checker model shares the generator's errors (risk of bias) and when the knowledge source is incomplete (indirectness). |
Related techniques
- should follow PE-12 Model summarisation of documents (abstractive summarisation; multi-document summarisation): long summaries are the usual output scored
- pairs with CK-16 Entailment-based consistency check (textual entailment; NLI (natural language inference) consistency; SummaC; FactCC; faithfulness checking): entailment scores sentences and this technique scores atomic facts
- pairs with CK-17 Source-span and quotation check by code (quote-and-verify; extractive grounding; citation verification): span checks confirm that supporting passages exist
Evidence and tag basis
Tested. The automated estimator had an error rate below 2 per cent against human scoring on the authors' biography task (2023, abstract). Do not transfer the figure to other tasks. (Min et al. 2023, arXiv 2305.14251)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.